No, I didn’t push any further after getting the Retina net and focal loss working…
You can search forum for object detection posts, I think last year people were talking about a lot of mAP and where their model was focusing on (heat maps).
For me, I only use it to understand how object detection model works. Also, there are lots of library such as Yolo, where you can use pretrained weights. I think that will be much easier to push SOTA (maybe I am wrong)
Also, you can check the on going kaggle google AI competition, people are talking a lot of object detection, segementation models.
Hope that helps